Sony Music and Warner Chappell Sue Anthropic Over Copyrighted Song Lyrics

Sony Music Publishing and Warner Chappell Music jointly filed a copyright infringement lawsuit against Anthropic on August 29, 2026, accusing the AI company of "one of the largest and most blatant ongoing thefts of intellectual property in history" (Engadget).

The complaint, filed on a Friday and published in full by Music Business Worldwide, alleges that Anthropic conducted a "brazen campaign of illegally torrenting, scraping and downloading copyrighted works on a massive scale" to train its Claude models (Engadget). Torrenting refers to using peer-to-peer file-sharing networks to download large quantities of data, while scraping means automated extraction of content from websites. The publishers claim that "thousands upon thousands of music publishers' copyrighted musical compositions" were illegally harvested, with damages that could reach into the billions (Engadget).

The lawsuit brings four counts against Anthropic, including direct infringement by torrenting and contributory infringement by torrenting (Music Business Worldwide). Direct infringement means the company itself participated in the unauthorized copying; contributory infringement means it knowingly helped or enabled others to do so. The publishers are seeking a jury trial, up to $150,000 for each infringed work, and $25,000 for each instance of Anthropic removing copyright management information (Engadget).

This is not the first time Anthropic has faced legal action from music rights holders. Earlier in 2026, Concord Music Group and Universal Music Group sued the company, seeking more than $3 billion in damages and accusing it of illegally downloading more than 20,000 copyrighted songs for AI training (Engadget). BMG filed its own lawsuit against Anthropic in March 2026, alleging the company trained Claude on unlicensed compositions by Bruno Mars and the Rolling Stones and that the chatbot now produces infringing lyrics (Billboard; Reuters).

The earliest of these actions dates to October 2023, when Universal Music Group sued Anthropic over "systematic and widespread infringement of their copyrighted song lyrics" by the Claude chatbot (CNBC). That same month, Reuters reported that the publishers' lawsuit alleged violations involving lyrics from at least 500 songs, ranging from the Beach Boys' "God Only Knows" to numerous other works (Reuters).

During the copyright litigation, Anthropic admitted that it copied the entirety of the publishers' copyrighted lyrics but asserted that it did not need to train on those lyrics (McKool Smith). Music publishers filed an amended lyrics lawsuit against Anthropic in July 2026, coinciding with court approval of Anthropic's separate $1.5 billion settlement with a group of authors who had accused the company of using pirated copies of their protected works for AI training (Music Business Worldwide). That settlement was described as record-setting (Engadget).

Anthropic's use of the publishers' lyrics to train its Claude generative AI model resulted in the unauthorized dissemination of those lyrics, according to legal analysis of the case (Lexology).

The broader music-versus-AI copyright landscape has seen parallel litigation and some resolution. In June 2024, Sony Music, Universal Music Group, and Warner Records sued AI music companies Suno and Udio in the US for copyright infringement (Reuters). Warner Music Group settled with both Udio and Suno in November 2025 on similar terms, and announced a joint platform with Udio (Reuters; Reuters). Sony Music's case against Suno remained ongoing as of March 2026 (Reuters).

The pattern across these cases is worth examining. Anthropic's $1.5 billion author settlement, the $3 billion-plus sought by Concord and UMG, and now the Sony-Warner Chappell complaint collectively signal that the financial exposure from training-data copyright claims has moved from theoretical to concrete and escalating. The Sony-Warner suit introduces the specific allegation of torrenting as a training-data acquisition method, which carries its own legal weight: direct infringement by torrenting and contributory infringement by torrenting are distinct counts that could establish a more aggressive posture than claims grounded solely in the output of trained models.

Also worth flagging is the admission in the existing litigation that Anthropic copied the entirety of the publishers' lyrics while simultaneously asserting it did not need to train on them. That combination is legally awkward for the defense. If the training was unnecessary, the copying becomes harder to justify under fair use, the legal doctrine that permits limited use of copyrighted material without permission, typically weighing the purpose and necessity of the use against its commercial impact. The publishers' claim that Claude now disseminates those lyrics compounds the exposure, because output-side infringement adds a separate vector of liability beyond the training-data ingestion itself.

The contrast with the Suno and Udio settlements is also relevant. Warner Music Group resolved those disputes and moved into a commercial partnership with Udio, suggesting that some rights holders see licensing as a viable path forward with AI music companies. The decision by Sony Music Publishing and Warner Chappell to escalate against Anthropic rather than settle indicates a different calculation, likely driven by the scale of alleged infringement, the torrenting allegations, and the fact that Anthropic is a general-purpose AI company rather than a music-specific one.

For technology professionals, the case underscores a practical point about training-data provenance that the industry has been circling for years. The methods used to assemble training corpora matter as much as the content itself. Allegations of torrenting, scraping, and removal of copyright management information describe specific technical actions that, if proven, establish a factual record difficult to reframe as incidental or transformative use. Companies building large language models now face a landscape where prior settlements have established billion-dollar baselines, multiple rights-holder groups are filing independently, and the legal system is beginning to draw lines between acceptable data acquisition and what the publishers describe as wholesale theft.